agora inbox for pgsql-hackers@postgresql.orghelp / color / mirror / Atom feed
[PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 268+ messages / 2 participants [nested] [flat]
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <lukas@fittl.com> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <geidav.pg@gmail.com> Author: Andres Freund <andres@anarazel.de> Author: Lukas Fittl <lukas@fittl.com> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v11 2/2] handle relation statistics correctly during rewrites @ 2025-11-04 13:52 Bertrand Drouvot <bertranddrouvot.pg@gmail.com> 0 siblings, 0 replies; 268+ messages in thread From: Bertrand Drouvot @ 2025-11-04 13:52 UTC (permalink / raw) Now that PGSTAT_KIND_RELATION is keyed by refilenode, we need to handle rewrites. To do so, this patch: - Adds PgStat_PendingRewrite, a new struct to track rewrite operations within a transaction, storing the old locator, new locator, and original locator (for rewrite chains). This allows stats to be copied from the original location to the final location at commit time. - Adds a new function, pgstat_mark_rewrite(), called when a table rewrite begins. It records the rewrite operation in a local list and detects rewrite chains by checking if the old_locator matches any existing new_locator, preserving the chain's original_locator. - Modifies pgstat_copy_relation_stats(), to accept RelFileLocators instead of Relations, with a new increment parameter to accumulate stats (needed for rewrite chains with DML between rewrites). - Ensures that AtEOXact_PgStat_Relations(), AtPrepare_PgStat_Relations(), pgstat_twophase_postcommit()/postabort() pgstat_drop_relation() handle the PgStat_PendingRewrite list correctly. Note that due to the new flush call in pgstat_twophase_postcommit() we can not call GetCurrentTransactionStopTimestamp() in pgstat_relation_flush_cb(). So, adding a check to handle this special case and call GetCurrentTimestamp() instead. Note that we'd call GetCurrentTimestamp() only if there is a rewrite, so that the GetCurrentTimestamp() extra cost should be negligible. Another solution could be to trigger the flush from FinishPreparedTransaction() but that's not worth the extra complexity. The new pending_rewrites list is traversed in multiple places. The overhead should be negligible in comparison to a rewrite and the list should not contain a lot of rewrites in practice. The pending_rewrites list is traversed in multiple places. In typical usage, the list will contain only a few entries so the traversal cost is negligible ( furthermore in comparison to a rewrite). --- src/backend/catalog/index.c | 2 +- src/backend/commands/cluster.c | 5 + src/backend/commands/tablecmds.c | 6 + src/backend/utils/activity/pgstat_relation.c | 391 ++++++++++++++++++- src/backend/utils/activity/pgstat_xact.c | 25 +- src/backend/utils/cache/relcache.c | 6 + src/include/pgstat.h | 5 +- src/tools/pgindent/typedefs.list | 1 + 8 files changed, 424 insertions(+), 17 deletions(-) 92.8% src/backend/utils/activity/ 4.9% src/backend/ diff --git a/src/backend/catalog/index.c b/src/backend/catalog/index.c index 43de42ce39e..88ef00609db 100644 --- a/src/backend/catalog/index.c +++ b/src/backend/catalog/index.c @@ -1793,7 +1793,7 @@ index_concurrently_swap(Oid newIndexId, Oid oldIndexId, const char *oldName) changeDependenciesOn(RelationRelationId, oldIndexId, newIndexId); /* copy over statistics from old to new index */ - pgstat_copy_relation_stats(newClassRel, oldClassRel); + pgstat_copy_relation_stats(newClassRel->rd_locator, oldClassRel->rd_locator, false); /* Copy data of pg_statistic from the old index to the new one */ CopyStatistics(oldIndexId, newIndexId); diff --git a/src/backend/commands/cluster.c b/src/backend/commands/cluster.c index 60a4617a585..3a446f32921 100644 --- a/src/backend/commands/cluster.c +++ b/src/backend/commands/cluster.c @@ -1196,6 +1196,11 @@ swap_relation_files(Oid r1, Oid r2, bool target_is_pg_class, rel1 = relation_open(r1, NoLock); rel2 = relation_open(r2, NoLock); + + /* Mark that a rewrite happened */ + if (RELKIND_HAS_STORAGE(rel1->rd_rel->relkind)) + pgstat_mark_rewrite(rel1->rd_locator, rel2->rd_locator); + rel2->rd_createSubid = rel1->rd_createSubid; rel2->rd_newRelfilelocatorSubid = rel1->rd_newRelfilelocatorSubid; rel2->rd_firstRelfilelocatorSubid = rel1->rd_firstRelfilelocatorSubid; diff --git a/src/backend/commands/tablecmds.c b/src/backend/commands/tablecmds.c index df1ba112b35..6244d04ab12 100644 --- a/src/backend/commands/tablecmds.c +++ b/src/backend/commands/tablecmds.c @@ -16896,6 +16896,7 @@ ATExecSetTableSpace(Oid tableOid, Oid newTableSpace, LOCKMODE lockmode) Oid reltoastrelid; RelFileNumber newrelfilenumber; RelFileLocator newrlocator; + RelFileLocator oldrlocator; List *reltoastidxids = NIL; ListCell *lc; @@ -16934,6 +16935,7 @@ ATExecSetTableSpace(Oid tableOid, Oid newTableSpace, LOCKMODE lockmode) newrlocator = rel->rd_locator; newrlocator.relNumber = newrelfilenumber; newrlocator.spcOid = newTableSpace; + oldrlocator = rel->rd_locator; /* hand off to AM to actually create new rel storage and copy the data */ if (rel->rd_rel->relkind == RELKIND_INDEX) @@ -16946,6 +16948,10 @@ ATExecSetTableSpace(Oid tableOid, Oid newTableSpace, LOCKMODE lockmode) table_relation_copy_data(rel, &newrlocator); } + /* mark that a rewrite happened */ + if (RELKIND_HAS_STORAGE(rel->rd_rel->relkind)) + pgstat_mark_rewrite(oldrlocator, newrlocator); + /* * Update the pg_class row. * diff --git a/src/backend/utils/activity/pgstat_relation.c b/src/backend/utils/activity/pgstat_relation.c index 89bf0cbed56..4d929972fab 100644 --- a/src/backend/utils/activity/pgstat_relation.c +++ b/src/backend/utils/activity/pgstat_relation.c @@ -30,6 +30,19 @@ #include "utils/syscache.h" #include "utils/timestamp.h" +/* Pending rewrite operations for stats copying */ +typedef struct PgStat_PendingRewrite +{ + RelFileLocator old_locator; + RelFileLocator new_locator; + RelFileLocator original_locator; + int nest_level; /* Transaction nesting level where rewrite + * occurred */ + struct PgStat_PendingRewrite *next; +} PgStat_PendingRewrite; + +/* The pending rewrites list for current transaction */ +static PgStat_PendingRewrite *pending_rewrites = NULL; /* Record that's written to 2PC state file when pgstat state is persisted */ typedef struct TwoPhasePgStatRecord @@ -43,6 +56,8 @@ typedef struct TwoPhasePgStatRecord PgStat_Counter deleted_pre_truncdrop; RelFileLocator locator; /* table's rd_locator */ bool truncdropped; /* was the relation truncated/dropped? */ + RelFileLocator rewrite_old_locator; + int rewrite_nest_level; } TwoPhasePgStatRecord; @@ -54,27 +69,70 @@ static void restore_truncdrop_counters(PgStat_TableXactStatus *trans); /* - * Copy stats between relations. This is used for things like REINDEX + * Copy stats between RelFileLocator. This is used for things like REINDEX * CONCURRENTLY. */ void -pgstat_copy_relation_stats(Relation dst, Relation src) +pgstat_copy_relation_stats(RelFileLocator dst, RelFileLocator src, bool increment) { PgStat_StatTabEntry *srcstats; PgStatShared_Relation *dstshstats; PgStat_EntryRef *dst_ref; - srcstats = pgstat_fetch_stat_tabentry_ext(RelationGetRelid(src)); + srcstats = (PgStat_StatTabEntry *) pgstat_fetch_entry(PGSTAT_KIND_RELATION, + src.dbOid, + RelFileLocatorToPgStatObjid(src)); if (!srcstats) return; dst_ref = pgstat_get_entry_ref_locked(PGSTAT_KIND_RELATION, - dst->rd_rel->relisshared ? InvalidOid : MyDatabaseId, - RelationGetRelid(dst), + dst.dbOid, + RelFileLocatorToPgStatObjid(dst), false); dstshstats = (PgStatShared_Relation *) dst_ref->shared_stats; - dstshstats->stats = *srcstats; + + if (!increment) + dstshstats->stats = *srcstats; + else + { + /* Increment those statistics */ +#define RELFSTAT_ACC(fld, stats_to_add) \ + (dstshstats->stats.fld += stats_to_add->fld) + RELFSTAT_ACC(numscans, srcstats); + RELFSTAT_ACC(tuples_returned, srcstats); + RELFSTAT_ACC(tuples_fetched, srcstats); + RELFSTAT_ACC(tuples_inserted, srcstats); + RELFSTAT_ACC(tuples_updated, srcstats); + RELFSTAT_ACC(tuples_deleted, srcstats); + RELFSTAT_ACC(tuples_hot_updated, srcstats); + RELFSTAT_ACC(tuples_newpage_updated, srcstats); + RELFSTAT_ACC(live_tuples, srcstats); + RELFSTAT_ACC(dead_tuples, srcstats); + RELFSTAT_ACC(mod_since_analyze, srcstats); + RELFSTAT_ACC(ins_since_vacuum, srcstats); + RELFSTAT_ACC(blocks_fetched, srcstats); + RELFSTAT_ACC(blocks_hit, srcstats); + RELFSTAT_ACC(vacuum_count, srcstats); + RELFSTAT_ACC(autovacuum_count, srcstats); + RELFSTAT_ACC(analyze_count, srcstats); + RELFSTAT_ACC(autoanalyze_count, srcstats); + RELFSTAT_ACC(total_vacuum_time, srcstats); + RELFSTAT_ACC(total_autovacuum_time, srcstats); + RELFSTAT_ACC(total_analyze_time, srcstats); + RELFSTAT_ACC(total_autoanalyze_time, srcstats); +#undef RELFSTAT_ACC + + /* Replace those statistics */ +#define RELFSTAT_REP(fld, stats_to_rep) \ + (dstshstats->stats.fld = stats_to_rep->fld) + RELFSTAT_REP(lastscan, srcstats); + RELFSTAT_REP(last_vacuum_time, srcstats); + RELFSTAT_REP(last_autovacuum_time, srcstats); + RELFSTAT_REP(last_analyze_time, srcstats); + RELFSTAT_REP(last_autoanalyze_time, srcstats); +#undef RELFSTAT_REP + } pgstat_unlock_entry(dst_ref); } @@ -136,6 +194,7 @@ void pgstat_assoc_relation(Relation rel) { RelFileLocator locator; + PgStat_TableStatus *pgstat_info; Assert(rel->pgstat_enabled); Assert(rel->pgstat_info == NULL); @@ -164,14 +223,54 @@ pgstat_assoc_relation(Relation rel) locator.relNumber = rel->rd_id; } + /* + * If this relation was rewritten during the current transaction we may be + * reopening it with its new RelFileLocator. In that case, continue using + * the stats entry associated with the old locator rather than creating a + * new one. This ensures all stats from before and after the rewrite are + * tracked in a single entry which will be properly copied to the new + * locator at transaction commit. + */ + if (pending_rewrites != NULL) + { + PgStat_PendingRewrite *rewrite; + + for (rewrite = pending_rewrites; rewrite != NULL; rewrite = rewrite->next) + { + if (locator.dbOid == rewrite->new_locator.dbOid && + locator.spcOid == rewrite->new_locator.spcOid && + locator.relNumber == rewrite->new_locator.relNumber) + { + pgstat_info = pgstat_prep_relation_pending(rewrite->old_locator); + goto found_entry; + } + } + } + /* Else find or make the PgStat_TableStatus entry, and update link */ - rel->pgstat_info = pgstat_prep_relation_pending(locator); + pgstat_info = pgstat_prep_relation_pending(locator); + +found_entry: + rel->pgstat_info = pgstat_info; + + /* + * For relations stats, we key by physical file location, not by relation + * OID. This means during operations like ALTER TYPE it's possible that + * the relation OID changes but the relfilenode stays the same (no actual + * rewrite needed). Unlink the old relation first. + */ + if (pgstat_info->relation != NULL && + pgstat_info->relation != rel) + { + pgstat_info->relation->pgstat_info = NULL; + pgstat_info->relation = NULL; + } /* don't allow link a stats to multiple relcache entries */ - Assert(rel->pgstat_info->relation == NULL); + Assert(pgstat_info->relation == NULL); /* mark this relation as the owner */ - rel->pgstat_info->relation = rel; + pgstat_info->relation = rel; } /* @@ -214,14 +313,37 @@ pgstat_drop_relation(Relation rel) { int nest_level = GetCurrentTransactionNestLevel(); PgStat_TableStatus *pgstat_info; + bool skip_transactional_drop = false; /* don't track stats for relations without storage */ if (!RELKIND_HAS_STORAGE(rel->rd_rel->relkind)) return; - pgstat_drop_transactional(PGSTAT_KIND_RELATION, - rel->rd_locator.dbOid, - RelFileLocatorToPgStatObjid(rel->rd_locator)); + /* Check if this drop is part of a pending rewrite */ + if (pending_rewrites != NULL) + { + PgStat_PendingRewrite *rewrite; + + for (rewrite = pending_rewrites; rewrite != NULL; rewrite = rewrite->next) + { + if (rel->rd_locator.dbOid == rewrite->old_locator.dbOid && + rel->rd_locator.spcOid == rewrite->old_locator.spcOid && + rel->rd_locator.relNumber == rewrite->old_locator.relNumber) + { + skip_transactional_drop = true; + break; + } + } + } + + /* + * If it is part of a rewrite, drop its stats later, for example in + * AtEOXact_PgStat_Relations(), so skip it here. + */ + if (!skip_transactional_drop) + pgstat_drop_transactional(PGSTAT_KIND_RELATION, + rel->rd_locator.dbOid, + RelFileLocatorToPgStatObjid(rel->rd_locator)); if (!pgstat_should_count_relation(rel)) return; @@ -666,6 +788,48 @@ AtEOXact_PgStat_Relations(PgStat_SubXactStatus *xact_state, bool isCommit) } tabstat->trans = NULL; } + + /* preserve the stats in case of rewrite */ + if (isCommit && pending_rewrites != NULL) + { + PgStat_PendingRewrite *rewrite; + PgStat_PendingRewrite *prev = NULL; + PgStat_PendingRewrite *current = pending_rewrites; + PgStat_PendingRewrite *next; + + /* reverse the rewrites list to process in chronological order */ + while (current != NULL) + { + next = current->next; + current->next = prev; + prev = current; + current = next; + } + + /* now process rewrites in chronological order */ + for (rewrite = prev; rewrite != NULL; rewrite = rewrite->next) + { + PgStat_EntryRef *old_entry_ref; + + old_entry_ref = pgstat_fetch_pending_entry(PGSTAT_KIND_RELATION, + rewrite->old_locator.dbOid, + RelFileLocatorToPgStatObjid(rewrite->old_locator)); + + if (old_entry_ref && old_entry_ref->pending) + pgstat_relation_flush_cb(old_entry_ref, false); + + pgstat_copy_relation_stats(rewrite->new_locator, + rewrite->old_locator, true); + + /* drop old locator's stats */ + if (!pgstat_drop_entry(PGSTAT_KIND_RELATION, + rewrite->old_locator.dbOid, + RelFileLocatorToPgStatObjid(rewrite->old_locator))) + pgstat_request_entry_refs_gc(); + } + } + + pending_rewrites = NULL; } /* @@ -681,6 +845,30 @@ AtEOSubXact_PgStat_Relations(PgStat_SubXactStatus *xact_state, bool isCommit, in PgStat_TableXactStatus *trans; PgStat_TableXactStatus *next_trans; + /* + * If we don't commit then remove the associated rewrites if any, to keep + * the rewrite chain in sync with what will be eventually committed. + */ + if (!isCommit) + { + PgStat_PendingRewrite **rewrite_ptr = &pending_rewrites; + + while (*rewrite_ptr != NULL) + { + if ((*rewrite_ptr)->nest_level >= nestDepth) + { + PgStat_PendingRewrite *to_remove = *rewrite_ptr; + + *rewrite_ptr = (*rewrite_ptr)->next; + pfree(to_remove); + } + else + { + rewrite_ptr = &((*rewrite_ptr)->next); + } + } + } + for (trans = xact_state->first; trans != NULL; trans = next_trans) { PgStat_TableStatus *tabstat; @@ -760,11 +948,19 @@ void AtPrepare_PgStat_Relations(PgStat_SubXactStatus *xact_state) { PgStat_TableXactStatus *trans; + PgStat_PendingRewrite *rewrite; + /* + * For each tabstat, find its matching rewrite and remove it from the + * pending rewrites list. This way, after processing all tabstats, pending + * rewrites will only contain rewrite only transactions. + */ for (trans = xact_state->first; trans != NULL; trans = trans->next) { PgStat_TableStatus *tabstat PG_USED_FOR_ASSERTS_ONLY; TwoPhasePgStatRecord record; + PgStat_PendingRewrite **rewrite_ptr; + bool found_rewrite = false; Assert(trans->nest_level == 1); Assert(trans->upper == NULL); @@ -784,10 +980,83 @@ AtPrepare_PgStat_Relations(PgStat_SubXactStatus *xact_state) record.locator = tabstat->locator; record.truncdropped = trans->truncdropped; + record.rewrite_nest_level = 0; + + /* + * Look for a matching rewrite and remove it from pending rewrites. We + * check three possible matches: + * + * The new_locator when stats have been added after the rewrite. The + * old_locator when stats have been added before the rewrite but not + * after. The original_locator when this tabstat is part of a rewrite + * chain. + */ + rewrite_ptr = &pending_rewrites; + while (*rewrite_ptr != NULL) + { + rewrite = *rewrite_ptr; + + if ((record.locator.dbOid == rewrite->new_locator.dbOid && + record.locator.spcOid == rewrite->new_locator.spcOid && + record.locator.relNumber == rewrite->new_locator.relNumber) || + (tabstat->locator.dbOid == rewrite->old_locator.dbOid && + tabstat->locator.spcOid == rewrite->old_locator.spcOid && + tabstat->locator.relNumber == rewrite->old_locator.relNumber) || + (tabstat->locator.dbOid == rewrite->original_locator.dbOid && + tabstat->locator.spcOid == rewrite->original_locator.spcOid && + tabstat->locator.relNumber == rewrite->original_locator.relNumber)) + { + /* + * Found matching rewrite. Record the rewrite information and + * remove this rewrite from the list since it's now handled. + */ + record.rewrite_old_locator = rewrite->original_locator; + record.rewrite_nest_level = rewrite->nest_level; + record.locator = rewrite->new_locator; + found_rewrite = true; + + /* Remove from pending_rewrites list */ + *rewrite_ptr = rewrite->next; + pfree(rewrite); + break; + } + else + { + /* Move to next rewrite in the list */ + rewrite_ptr = &(rewrite->next); + } + } + + /* If no rewrite found, clear the rewrite fields */ + if (!found_rewrite) + { + memset(&record.rewrite_old_locator, 0, sizeof(RelFileLocator)); + } + + RegisterTwoPhaseRecord(TWOPHASE_RM_PGSTAT_ID, 0, + &record, sizeof(TwoPhasePgStatRecord)); + } + + /* + * Now process any rewrites still pending. These are rewrite only + * transactions. We need to preserve their stats even though there's no + * tabstat entry for them. + */ + for (rewrite = pending_rewrites; rewrite != NULL; rewrite = rewrite->next) + { + TwoPhasePgStatRecord record; + + memset(&record, 0, sizeof(TwoPhasePgStatRecord)); + record.locator = rewrite->new_locator; + record.rewrite_old_locator = rewrite->original_locator; + record.rewrite_nest_level = rewrite->nest_level; + record.truncdropped = false; RegisterTwoPhaseRecord(TWOPHASE_RM_PGSTAT_ID, 0, &record, sizeof(TwoPhasePgStatRecord)); } + + pending_rewrites = NULL; } /* @@ -810,6 +1079,8 @@ PostPrepare_PgStat_Relations(PgStat_SubXactStatus *xact_state) tabstat = trans->parent; tabstat->trans = NULL; } + + pending_rewrites = NULL; } /* @@ -845,6 +1116,29 @@ pgstat_twophase_postcommit(FullTransactionId fxid, uint16 info, pgstat_info->counts.changed_tuples += rec->tuples_inserted + rec->tuples_updated + rec->tuples_deleted; + + if (rec->rewrite_nest_level > 0) + { + PgStat_EntryRef *old_entry_ref; + + /* Flush any pending stats for old locator first */ + old_entry_ref = pgstat_fetch_pending_entry(PGSTAT_KIND_RELATION, + rec->rewrite_old_locator.dbOid, + RelFileLocatorToPgStatObjid(rec->rewrite_old_locator)); + + if (old_entry_ref && old_entry_ref->pending) + pgstat_relation_flush_cb(old_entry_ref, false); + + /* Copy stats from old to new locator */ + pgstat_copy_relation_stats(rec->locator, rec->rewrite_old_locator, + true); + + /* Drop old locator's stats */ + if (!pgstat_drop_entry(PGSTAT_KIND_RELATION, + rec->rewrite_old_locator.dbOid, + RelFileLocatorToPgStatObjid(rec->rewrite_old_locator))) + pgstat_request_entry_refs_gc(); + } } /* @@ -859,9 +1153,26 @@ pgstat_twophase_postabort(FullTransactionId fxid, uint16 info, { TwoPhasePgStatRecord *rec = (TwoPhasePgStatRecord *) recdata; PgStat_TableStatus *pgstat_info; + RelFileLocator target_locator; + + /* + * For aborted transactions with rewrites (like TRUNCATE), we need to + * restore stats to the old locator, not the new one. The new locator + * should be dropped since the rewrite is being rolled back. + */ + if (rec->rewrite_nest_level > 0) + { + /* Use the old locator */ + target_locator = rec->rewrite_old_locator; + } + else + { + /* No rewrite, use the original locator */ + target_locator = rec->locator; + } /* Find or create a tabstat entry for the target locator */ - pgstat_info = pgstat_prep_relation_pending(rec->locator); + pgstat_info = pgstat_prep_relation_pending(target_locator); /* Same math as in AtEOXact_PgStat, abort case */ if (rec->truncdropped) @@ -916,7 +1227,17 @@ pgstat_relation_flush_cb(PgStat_EntryRef *entry_ref, bool nowait) tabentry->numscans += lstats->counts.numscans; if (lstats->counts.numscans) { - TimestampTz t = GetCurrentTransactionStopTimestamp(); + TimestampTz t; + + /* + * Checking the transaction state due to the flush call in + * pgstat_twophase_postcommit() that would break the assertion on the + * state in GetCurrentTransactionStopTimestamp(). + */ + if (!IsTransactionState()) + t = GetCurrentTransactionStopTimestamp(); + else + t = GetCurrentTimestamp(); if (t > tabentry->lastscan) tabentry->lastscan = t; @@ -1167,3 +1488,45 @@ pgstat_reloid_to_relfilelocator(Oid reloid, RelFileLocator *locator) ReleaseSysCache(tuple); return result; } + +/* + * Mark that a relation rewrite has occurred, preserving the original locator + * so stats can be copied at transaction commit. + */ +void +pgstat_mark_rewrite(RelFileLocator old_locator, RelFileLocator new_locator) +{ + PgStat_PendingRewrite *rewrite; + PgStat_PendingRewrite *existing; + RelFileLocator original_locator = old_locator; + + for (existing = pending_rewrites; existing != NULL; existing = existing->next) + { + if (old_locator.dbOid == existing->new_locator.dbOid && + old_locator.spcOid == existing->new_locator.spcOid && + old_locator.relNumber == existing->new_locator.relNumber) + { + original_locator = existing->original_locator; + break; + } + } + + /* Allocate in TopTransactionContext memory context */ + rewrite = MemoryContextAlloc(TopTransactionContext, + sizeof(PgStat_PendingRewrite)); + + rewrite->old_locator = old_locator; + rewrite->new_locator = new_locator; + rewrite->original_locator = original_locator; + rewrite->nest_level = GetCurrentTransactionNestLevel(); + + /* Add to the list */ + rewrite->next = pending_rewrites; + pending_rewrites = rewrite; +} + +void +pgstat_clear_rewrite(void) +{ + pending_rewrites = NULL; +} diff --git a/src/backend/utils/activity/pgstat_xact.c b/src/backend/utils/activity/pgstat_xact.c index 5e2d69e6297..8ed8f5317f3 100644 --- a/src/backend/utils/activity/pgstat_xact.c +++ b/src/backend/utils/activity/pgstat_xact.c @@ -55,6 +55,8 @@ AtEOXact_PgStat(bool isCommit, bool parallel) } pgStatXactStack = NULL; + pgstat_clear_rewrite(); + /* Make sure any stats snapshot is thrown away */ pgstat_clear_snapshot(); } @@ -360,8 +362,29 @@ create_drop_transactional_internal(PgStat_Kind kind, Oid dboid, uint64 objid, bo void pgstat_create_transactional(PgStat_Kind kind, Oid dboid, uint64 objid) { - if (pgstat_get_entry_ref(kind, dboid, objid, false, NULL)) + PgStat_EntryRef *entry_ref; + + entry_ref = pgstat_get_entry_ref(kind, dboid, objid, false, NULL); + + if (entry_ref) { + /* + * For relations stats, we key by physical file location, not by + * relation OID. This means during operations like ALTER TYPE where + * the relation OID changes but the relfilenode stays the same (no + * actual rewrite needed), we'll find an existing entry. + * + * This is expected behavior, we want to preserve stats across the + * catalog change. Simply reset and recreate the entry for the new + * relation OID without warning. + */ + if (kind == PGSTAT_KIND_RELATION) + { + pgstat_reset(kind, dboid, objid); + create_drop_transactional_internal(kind, dboid, objid, true); + return; + } + ereport(WARNING, errmsg("resetting existing statistics for kind %s, db=%u, oid=%" PRIu64, (pgstat_get_kind_info(kind))->name, dboid, diff --git a/src/backend/utils/cache/relcache.c b/src/backend/utils/cache/relcache.c index 6b634c9fff1..692d084acb2 100644 --- a/src/backend/utils/cache/relcache.c +++ b/src/backend/utils/cache/relcache.c @@ -85,6 +85,7 @@ #include "utils/inval.h" #include "utils/lsyscache.h" #include "utils/memutils.h" +#include "utils/pgstat_internal.h" #include "utils/relmapper.h" #include "utils/resowner.h" #include "utils/snapmgr.h" @@ -3775,6 +3776,7 @@ RelationSetNewRelfilenumber(Relation relation, char persistence) MultiXactId minmulti = InvalidMultiXactId; TransactionId freezeXid = InvalidTransactionId; RelFileLocator newrlocator; + RelFileLocator oldrlocator = relation->rd_locator; if (!IsBinaryUpgrade) { @@ -3946,6 +3948,10 @@ RelationSetNewRelfilenumber(Relation relation, char persistence) table_close(pg_class, RowExclusiveLock); + /* Mark that a rewrite happened */ + if (RELKIND_HAS_STORAGE(relation->rd_rel->relkind)) + pgstat_mark_rewrite(oldrlocator, newrlocator); + /* * Make the pg_class row change or relation map change visible. This will * cause the relcache entry to get updated, too. diff --git a/src/include/pgstat.h b/src/include/pgstat.h index e9e4d32c3b8..a9a4fa9f8f2 100644 --- a/src/include/pgstat.h +++ b/src/include/pgstat.h @@ -669,7 +669,7 @@ extern PgStat_FunctionCounts *find_funcstat_entry(Oid func_id); extern void pgstat_create_relation(Relation rel); extern void pgstat_drop_relation(Relation rel); -extern void pgstat_copy_relation_stats(Relation dst, Relation src); +extern void pgstat_copy_relation_stats(RelFileLocator dst, RelFileLocator src, bool increment); extern void pgstat_init_relation(Relation rel); extern void pgstat_assoc_relation(Relation rel); @@ -681,6 +681,9 @@ extern void pgstat_report_vacuum(Relation rel, PgStat_Counter livetuples, extern void pgstat_report_analyze(Relation rel, PgStat_Counter livetuples, PgStat_Counter deadtuples, bool resetcounter, TimestampTz starttime); +extern void pgstat_mark_rewrite(RelFileLocator old_locator, + RelFileLocator new_locator); +extern void pgstat_clear_rewrite(void); /* * If stats are enabled, but pending data hasn't been prepared yet, call diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list index 241945734ec..05b1380c08c 100644 --- a/src/tools/pgindent/typedefs.list +++ b/src/tools/pgindent/typedefs.list @@ -2279,6 +2279,7 @@ PgStat_KindInfo PgStat_LocalState PgStat_PendingDroppedStatsItem PgStat_PendingIO +PgStat_PendingRewrite PgStat_SLRUStats PgStat_ShmemControl PgStat_Snapshot -- 2.34.1 --zLA8DXafBEP889wv-- ^ permalink raw reply [nested|flat] 268+ messages in thread
end of thread, other threads:[~2025-11-04 13:52 UTC | newest] Thread overview: 268+ messages (download: mbox mbox.gz follow: Atom feed) -- links below jump to the message on this page -- 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <lukas@fittl.com> 2025-11-04 13:52 [PATCH v11 2/2] handle relation statistics correctly during rewrites Bertrand Drouvot <bertranddrouvot.pg@gmail.com>
This inbox is served by agora; see mirroring instructions for how to clone and mirror all data and code used for this inbox