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 v4 3/3] Adding the pg_stat_backend_transaction view @ 2025-08-09 14:22 Bertrand Drouvot <bertranddrouvot.pg@gmail.com> 0 siblings, 0 replies; 268+ messages in thread From: Bertrand Drouvot @ 2025-08-09 14:22 UTC (permalink / raw) This view displays one row per server process, showing transaction statistics related to the current activity of that process. It currently displays the pid, the number of XIDs generated, the number of commits, the number of rollbacks and the time at which these statistics were last reset. It's built on top of a new function (pg_stat_get_backend_transactions()). The idea is the same as pg_stat_activity and pg_stat_get_activity(). Adding documentation and tests. XXX: Bump catversion --- doc/src/sgml/monitoring.sgml | 115 +++++++++++++++++++++++++++ src/backend/catalog/system_views.sql | 9 +++ src/backend/utils/adt/pgstatfuncs.c | 57 +++++++++++++ src/include/catalog/pg_proc.dat | 9 +++ src/test/regress/expected/rules.out | 6 ++ src/test/regress/expected/stats.out | 17 ++++ src/test/regress/sql/stats.sql | 10 +++ 7 files changed, 223 insertions(+) 51.7% doc/src/sgml/ 3.2% src/backend/catalog/ 23.1% src/backend/utils/adt/ 7.0% src/include/catalog/ 9.6% src/test/regress/expected/ 5.1% src/test/regress/sql/ diff --git a/doc/src/sgml/monitoring.sgml b/doc/src/sgml/monitoring.sgml index 3f4a27a736e..025641c2f70 100644 --- a/doc/src/sgml/monitoring.sgml +++ b/doc/src/sgml/monitoring.sgml @@ -320,6 +320,20 @@ postgres 27093 0.0 0.0 30096 2752 ? Ss 11:34 0:00 postgres: ser </entry> </row> + <row> + <entry> + <structname>pg_stat_backend_transaction</structname> + <indexterm><primary>pg_stat_backend_transaction</primary></indexterm> + </entry> + <entry> + One row per server process, showing statistics related to + the current activity of that process, such as number of commits and + rollbacks. + See <link linkend="monitoring-pg-stat-backend-transaction-view"> + <structname>pg_stat_backend_transaction</structname></link> for details. + </entry> + </row> + <row> <entry><structname>pg_stat_replication</structname><indexterm><primary>pg_stat_replication</primary></indexterm></entry> <entry>One row per WAL sender process, showing statistics about @@ -1172,6 +1186,91 @@ description | Waiting for a newly initialized WAL file to reach durable storage </note> </sect2> + <sect2 id="monitoring-pg-stat-backend-transaction-view"> + <title><structname>pg_stat_backend_transaction</structname></title> + + <indexterm> + <primary>pg_stat_backend_transaction</primary> + </indexterm> + + <para> + The <structname>pg_stat_backend_transaction</structname> view will have one row + per server process, showing statistics related to + the current activity of that process. + </para> + + <table id="pg-stat-backend-transaction-view" xreflabel="pg_stat_backend_transaction"> + <title><structname>pg_stat_backend_transaction</structname> View</title> + <tgroup cols="1"> + <thead> + <row> + <entry role="catalog_table_entry"><para role="column_definition"> + Column Type + </para> + <para> + Description + </para></entry> + </row> + </thead> + + <tbody> + <row> + <entry role="catalog_table_entry"><para role="column_definition"> + <structfield>pid</structfield> <type>integer</type> + </para> + <para> + Process ID of this backend + </para></entry> + </row> + + <row> + <entry role="catalog_table_entry"><para role="column_definition"> + <structfield>xid_count</structfield> <type>bigint</type> + </para> + <para> + The number of XID that have been generated by the backend. It does not take + into account virtual transaction ID on purpose. + </para></entry> + </row> + + <row> + <entry role="catalog_table_entry"><para role="column_definition"> + <structfield>xact_commit</structfield> <type>bigint</type> + </para> + <para> + The number of transactions that have been committed. + </para></entry> + </row> + + <row> + <entry role="catalog_table_entry"><para role="column_definition"> + <structfield>xact_rollback</structfield> <type>bigint</type> + </para> + <para> + The number of transactions that have been rolled back. + </para></entry> + </row> + + <row> + <entry role="catalog_table_entry"><para role="column_definition"> + <structfield>stats_reset</structfield> <type>timestamp with time zone</type> + </para> + <para> + Time at which these statistics were last reset + </para></entry> + </row> + </tbody> + </tgroup> + </table> + + <note> + <para> + The view does not return statistics for the checkpointer, + the background writer, the startup process and the autovacuum launcher. + </para> + </note> + </sect2> + <sect2 id="monitoring-pg-stat-replication-view"> <title><structname>pg_stat_replication</structname></title> @@ -4921,6 +5020,22 @@ description | Waiting for a newly initialized WAL file to reach durable storage </para></entry> </row> + <row> + <entry role="func_table_entry"><para role="func_signature"> + <indexterm> + <primary>pg_stat_get_backend_transactions</primary> + </indexterm> + <function>pg_stat_get_backend_transactions</function> ( <type>integer</type> ) + <returnvalue>setof record</returnvalue> + </para> + <para> + Returns a record of transaction statistics about the backend with the + specified process ID, or one record for each active backend in the system + if <literal>NULL</literal> is specified. The fields returned are a + subset of those in the <structname>pg_stat_backend_transaction</structname> view. + </para></entry> + </row> + <row> <entry id="pg-stat-get-backend-wal" role="func_table_entry"><para role="func_signature"> <indexterm> diff --git a/src/backend/catalog/system_views.sql b/src/backend/catalog/system_views.sql index 1b3c5a55882..ee3f653f360 100644 --- a/src/backend/catalog/system_views.sql +++ b/src/backend/catalog/system_views.sql @@ -911,6 +911,15 @@ CREATE VIEW pg_stat_activity AS LEFT JOIN pg_database AS D ON (S.datid = D.oid) LEFT JOIN pg_authid AS U ON (S.usesysid = U.oid); +CREATE VIEW pg_stat_backend_transaction AS + SELECT + S.pid, + S.xid_count, + S.xact_commit, + S.xact_rollback, + S.stats_reset + FROM pg_stat_get_backend_transactions(NULL) AS S; + CREATE VIEW pg_stat_replication AS SELECT S.pid, diff --git a/src/backend/utils/adt/pgstatfuncs.c b/src/backend/utils/adt/pgstatfuncs.c index c756c2bebaa..fa92099e789 100644 --- a/src/backend/utils/adt/pgstatfuncs.c +++ b/src/backend/utils/adt/pgstatfuncs.c @@ -685,6 +685,63 @@ pg_stat_get_activity(PG_FUNCTION_ARGS) return (Datum) 0; } +/* + * Returns statistics of PG backends. + */ +Datum +pg_stat_get_backend_transactions(PG_FUNCTION_ARGS) +{ +#define PG_STAT_GET_BACKEND_STATS_COLS 5 + int num_backends = pgstat_fetch_stat_numbackends(); + int curr_backend; + int pid = PG_ARGISNULL(0) ? -1 : PG_GETARG_INT32(0); + ReturnSetInfo *rsinfo = (ReturnSetInfo *) fcinfo->resultinfo; + + InitMaterializedSRF(fcinfo, 0); + + /* 1-based index */ + for (curr_backend = 1; curr_backend <= num_backends; curr_backend++) + { + /* for each row */ + Datum values[PG_STAT_GET_BACKEND_STATS_COLS] = {0}; + bool nulls[PG_STAT_GET_BACKEND_STATS_COLS] = {0}; + LocalPgBackendStatus *local_beentry; + PgBackendStatus *beentry; + PgStat_Backend *backend_stats; + + /* Get the next one in the list */ + local_beentry = pgstat_get_local_beentry_by_index(curr_backend); + beentry = &local_beentry->backendStatus; + + /* If looking for specific PID, ignore all the others */ + if (pid != -1 && beentry->st_procpid != pid) + continue; + + backend_stats = pgstat_fetch_stat_backend_by_pid(beentry->st_procpid, NULL); + + values[0] = Int32GetDatum(beentry->st_procpid); + + if (!backend_stats) + continue; + + values[1] = Int64GetDatum(backend_stats->xid_count); + values[2] = Int64GetDatum(backend_stats->xact_commit); + values[3] = Int64GetDatum(backend_stats->xact_rollback); + + if (backend_stats->stat_reset_timestamp != 0) + values[4] = TimestampTzGetDatum(backend_stats->stat_reset_timestamp); + else + nulls[4] = true; + + tuplestore_putvalues(rsinfo->setResult, rsinfo->setDesc, values, nulls); + + /* If only a single backend was requested, and we found it, break. */ + if (pid != -1) + break; + } + + return (Datum) 0; +} Datum pg_backend_pid(PG_FUNCTION_ARGS) diff --git a/src/include/catalog/pg_proc.dat b/src/include/catalog/pg_proc.dat index 118d6da1ace..24ec3d861a2 100644 --- a/src/include/catalog/pg_proc.dat +++ b/src/include/catalog/pg_proc.dat @@ -5641,6 +5641,15 @@ proargmodes => '{i,o,o,o,o,o,o,o,o,o,o,o,o,o,o,o,o,o,o,o,o,o,o,o,o,o,o,o,o,o,o,o}', proargnames => '{pid,datid,pid,usesysid,application_name,state,query,wait_event_type,wait_event,xact_start,query_start,backend_start,state_change,client_addr,client_hostname,client_port,backend_xid,backend_xmin,backend_type,ssl,sslversion,sslcipher,sslbits,ssl_client_dn,ssl_client_serial,ssl_issuer_dn,gss_auth,gss_princ,gss_enc,gss_delegation,leader_pid,query_id}', prosrc => 'pg_stat_get_activity' }, +{ oid => '9555', + descr => 'statistics: statistics about currently active backends', + proname => 'pg_stat_get_backend_transactions', prorows => '100', proisstrict => 'f', + proretset => 't', provolatile => 's', proparallel => 'r', + prorettype => 'record', proargtypes => 'int4', + proallargtypes => '{int4,int4,int8,int8,int8,timestamptz}', + proargmodes => '{i,o,o,o,o,o}', + proargnames => '{pid,pid,xid_count,xact_commit,xact_rollback,stats_reset}', + prosrc => 'pg_stat_get_backend_transactions' }, { oid => '6318', descr => 'describe wait events', proname => 'pg_get_wait_events', procost => '10', prorows => '250', proretset => 't', provolatile => 'v', prorettype => 'record', diff --git a/src/test/regress/expected/rules.out b/src/test/regress/expected/rules.out index 35e8aad7701..ae376f96998 100644 --- a/src/test/regress/expected/rules.out +++ b/src/test/regress/expected/rules.out @@ -1847,6 +1847,12 @@ pg_stat_archiver| SELECT archived_count, last_failed_time, stats_reset FROM pg_stat_get_archiver() s(archived_count, last_archived_wal, last_archived_time, failed_count, last_failed_wal, last_failed_time, stats_reset); +pg_stat_backend_transaction| SELECT pid, + xid_count, + xact_commit, + xact_rollback, + stats_reset + FROM pg_stat_get_backend_transactions(NULL::integer) s(pid, xid_count, xact_commit, xact_rollback, stats_reset); pg_stat_bgwriter| SELECT pg_stat_get_bgwriter_buf_written_clean() AS buffers_clean, pg_stat_get_bgwriter_maxwritten_clean() AS maxwritten_clean, pg_stat_get_buf_alloc() AS buffers_alloc, diff --git a/src/test/regress/expected/stats.out b/src/test/regress/expected/stats.out index 605f5070376..7079198bc06 100644 --- a/src/test/regress/expected/stats.out +++ b/src/test/regress/expected/stats.out @@ -135,11 +135,28 @@ INSERT INTO trunc_stats_test1 DEFAULT VALUES; INSERT INTO trunc_stats_test1 DEFAULT VALUES; UPDATE trunc_stats_test1 SET id = id + 10 WHERE id IN (1, 2); DELETE FROM trunc_stats_test1 WHERE id = 3; +-- in passing, check that backend's commit is incrementing +SELECT xact_commit AS xact_commit_before + FROM pg_stat_backend_transaction WHERE pid = pg_backend_pid() \gset BEGIN; UPDATE trunc_stats_test1 SET id = id + 100; TRUNCATE trunc_stats_test1; INSERT INTO trunc_stats_test1 DEFAULT VALUES; COMMIT; +SELECT pg_stat_force_next_flush(); + pg_stat_force_next_flush +-------------------------- + +(1 row) + +SELECT xact_commit AS xact_commit_after + FROM pg_stat_backend_transaction WHERE pid = pg_backend_pid() \gset +SELECT :xact_commit_after > :xact_commit_before; + ?column? +---------- + t +(1 row) + -- use a savepoint: 1 insert, 1 live BEGIN; INSERT INTO trunc_stats_test2 DEFAULT VALUES; diff --git a/src/test/regress/sql/stats.sql b/src/test/regress/sql/stats.sql index 54e72866344..0d6a408c7b5 100644 --- a/src/test/regress/sql/stats.sql +++ b/src/test/regress/sql/stats.sql @@ -58,12 +58,22 @@ INSERT INTO trunc_stats_test1 DEFAULT VALUES; UPDATE trunc_stats_test1 SET id = id + 10 WHERE id IN (1, 2); DELETE FROM trunc_stats_test1 WHERE id = 3; +-- in passing, check that backend's commit is incrementing +SELECT xact_commit AS xact_commit_before + FROM pg_stat_backend_transaction WHERE pid = pg_backend_pid() \gset + BEGIN; UPDATE trunc_stats_test1 SET id = id + 100; TRUNCATE trunc_stats_test1; INSERT INTO trunc_stats_test1 DEFAULT VALUES; COMMIT; +SELECT pg_stat_force_next_flush(); +SELECT xact_commit AS xact_commit_after + FROM pg_stat_backend_transaction WHERE pid = pg_backend_pid() \gset + +SELECT :xact_commit_after > :xact_commit_before; + -- use a savepoint: 1 insert, 1 live BEGIN; INSERT INTO trunc_stats_test2 DEFAULT VALUES; -- 2.34.1 --jg9kqX3osLGO+8A7-- ^ permalink raw reply [nested|flat] 268+ messages in thread
end of thread, other threads:[~2025-08-09 14:22 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> 2025-08-09 14:22 [PATCH v4 3/3] Adding the pg_stat_backend_transaction view 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